ORIGINAL PAPER

Evaluation of the use of low-density LiDAR data to estimatestructural attributes and biomass yield in a short-rotation willowcoppice: an example in a field trial

María Castaño-Díaz1 & Pedro Álvarez-Álvarez1 & Brian Tobin2&

Maarten Nieuwenhuis2 & Elías Afif-Khouri1 & Asunción Cámara-Obregón1

Received: 16 February 2017 /Accepted: 28 September 2017 /Published online: 23 October 2017# INRA and Springer-Verlag France SAS 2017

Abstract& Key message LiDAR data (low-density data, 0.5 pulsesm−2) represent an excellent management resource as theycan be used to estimate forest stand characteristics inshort-rotation willow coppice (SRWC) with reason-able accuracy. The technology is also a useful, prac-tical tool for carrying out inventories in these typesof stands.& Context This study evaluated the use of very low-densityairborne LiDAR (light detection and ranging) data (0.5 pulses

m−2), which can be accessed free of charge, in an SRWCestablished in degraded mining land.& Aims This work aimed to determine the utility of low-density LiDAR data for estimating main forest structural attri-butes and biomass productivity and for comparing the esti-mates with field measurements carried out in an SRWCplanted in marginal land.& Methods The SRWC was established following a random-ized complete block design with three clones, planted at twodensities and with three fertilization levels. Use of para-metric (multiple regression) and non-parametric (classi-fication and regression trees, CART) fitting techniquesyielded models with good predictive power and reliability.Both fitting methods were used for comprehensive analysisof the data and provide complementary information.& Results The results of multiple regression analysis indicatedclose relationships (Rfit

2 = 0.63–0.97) between LiDAR-derived metrics and the field measured data for the variablesstudied (H, D20, D130, FW, and DW). High R2 values wereobtained for models fitted using the CART technique(R2 = 0.73–0.94).& Conclusion Low-density LiDAR data can be used to modelstructural attributes and biomass yield in SRWC with reason-

* María Castaño-Díazcastanodmaria.uo@uniovi.es

Pedro Álvarez-Álvarezalvarezpedro@uniovi.es

Brian Tobinbrian.tobin@ucd.ie

Maarten Nieuwenhuismaarten.nieuwenhuis@ucd.ie

Elías Afif-Khourielias@uniovi.es

Asunción Cámara-Obregóncamara@uniovi.es

1 GIS-Forest Research Group, Department of Organismand Systems Biology, University of Oviedo,E-33600 Mieres, Spain

2 UCD Forestry, Agriculture and Food Science Centre,University College Dublin, Belfield, Dublin 4, Ireland

Contribution of the co-authors- María CASTAÑO-DÍAZ = co-writing the paper and analyzing the data.- Pedro ÁLVAREZ-ÁLVAREZ = co-writing the paper and analyzing thedata.- Brian TOBIN = co-writing the paper and supervising the work.- Maarten NIEUWENHUIS = co-writing the paper and supervising thework.- Elías AFIF-KHOURI = supervising the work.- Asunción CÁMARA-OBREGÓN = supervising the work and coordi-nating the research project.

Handling Editor: Aaron R. Weiskittel

Electronic supplementary materialThe online version of this article(https://doi.org/10.1007/s13595-017-0665-7)) contains supplementarymaterial, which is available to authorized users.

Annals of Forest Science (2017) 74: 69https://doi.org/10.1007/s13595-017-0665-7

able accuracy. The models developed can be used to improveand optimize follow-up decisions about the management ofthese crops.

Keywords Willow .Mining land . Energy crops . SRC .

Airborne laser scanning . LiDAR

1 Introduction

The prospects of successfully achieving and maintaining sus-tainable energy production worldwide depend on the in-creased use of renewable resources in general and biomassin particular (Edenhofer et al. 2011). One of the best ways ofensuring the long-term availability of biomass for producingrenewable energy is to establish and grow new perennial en-ergy crops, which can also add value to marginal land (Rossoet al. 2013) or can be used for bioremediation purposes.

Biomass plantations are an attractive source of renewableenergy (González-Ferreiro et al. 2013) and also have manyother advantages. Some of the reasons why crops are grownfor bioenergy purposes include the recovery of economic ac-tivities in rural areas, provision of a neutral CO2 balance, andrestoration of degraded land.

Depending on the final destination (heat and/or electricity),three types of biomass can be produced: oilseed, alcohol, andlignocellulose (IDAE 2007). Crops that produce lignocellu-losic biomass (fiber crops) can be used to produce both heatand electricity and can be grown as short-rotation coppice(SRC). Producing these so-called energy crops is consideredone of the most energy-efficient methods of carbon conver-sion, as growing the crops is considered an efficient means ofreducing greenhouse gas emissions (Styles and Jones 2007).

Short-rotation plantations can be established on varioustypes of land, including marginal land (Broeckx et al. 2012).Zurba et al. (2013) recommended planting SRC on marginalland and brownfields, in parallel with other sustainable landmanagement options. Planting SRC on this type of land mayalso contribute in the long term to improving soil quality andbiodiversity, protecting groundwater and preventing soil ero-sion (Kuzovkina and Quigley 2005; Zurba et al. 2013). Theuse of Salicaceae (Salix and Populus spp.) provides severaladvantages: the ease of propagating plants from cuttings (lowproduction cost and easy to establish), the wide range of im-proved genetic material available, production of high biomassyields in a short time, and vigorous coppicing regrowth aftercutting (Keoleian and Volk 2005). Taking into account the wideadaptability of members of the genus Salix to extreme condi-tions and to nutrient-impoverished and polluted soils(Kuzovkina and Quigley 2005), SRCwillow can be establishedon marginal land or in soils that are not suitable for agriculturalexploitation (Jama and Nowak 2012). Indeed, short-rotation

willow coppice (SRWC), together with Populus spp.,Eucalyptus spp., and Robinia pseudoacacia L, is one of themost promising bioenergy cropping systems for use in temper-ate regions of Europe (Venturi et al. 1999) as well as in Canadaand the USA (Tahvanainen and Rytko 1999; Weih 2004).

The region of Asturias (north-western Spain) was a majorcoal-producing region during the past century. Although coalmining continues to be one of the most important sources ofemployment in the region (Paredes-Sánchez et al. 2016;Suárez-Antuña 2005), the sector is currently in recession,and large areas of mining land have been abandoned. Themining company Grupo Hunosa currently owns up to700 ha of former mining land that is suitable for machine-based establishment of forest energy crops. This is currentlyconsidered the best option for use of this land, despite thedifficulties in establishing energy crops (unfavorable soilstructure/properties in these degraded areas). To date, the onlytrials involving SRC energy crops in Asturias are those asso-ciated with research projects (7 ha). At present, a commercialplantation (for bioenergy purposes) is being established in20 ha of abandoned mine land of similar characteristics tothose considered in the present study.

In 2008, an experimental trial with willow energy cropswas established in abandoned mining land in Asturias. Theaims of this experimental trial were to obtain informationabout structural attributes and biomass production in anSRWC crop established in a restored coal mining area and toevaluate the effects of clone, fertilization, and planting densityon crop yield. For this purpose, detailed and comprehensivefield inventories were conducted in order to obtain as muchinformation as possible about the development of the energycrop. Forest inventories were used to estimate multiple param-eters at plot level, including structure and biomass production.

For some decades now, remote sensing has enabledinformation about forest biomass (particularly in exten-sive forest areas) to be obtained at a wide range ofspatial and temporal scales, thus greatly reducing costsand the amount of fieldwork required (MontealegreGracia et al. 2015). The correlation between the spectralresponse of vegetation and structural attributes or bio-mass production has been investigated in numerousstudies in which active sensors were used (Estornellet al. 2011; González-Ferreiro et al. 2012; Næsset2002; Næsset and Gobakken 2008).

The interest shown by the aforementioned company indeveloping and applying new procedures and the possibil-ity of obtaining data (forest structure and other forest var-iables) from airborne sensors provides a valuable opportu-nity to quantify the resources obtained directly from SRC.This represents a breakthrough in this field, as carrying outthe field inventories necessary for adequate planning andmonitoring of the forest energy plantations (characterizedby high tree densities of 5000–20,000 plants ha−1, high

69 Page 2 of 16 Annals of Forest Science (2017) 74: 69

number of shoots per stool, etc.) is tedious and timeconsuming.

The use of free light detection and ranging (LiDAR)-de-rived data provides certain advantages such as smaller estima-tion errors, a reduction in the duration of field inventories andthe ability to cover larger areas of land.

LIDAR is one of the most important technologies developedin this field in recent years. This technique is already being usedsuccessfully to evaluate total forest area, improving the accuracyof forest inventories and reducing the cost and time spent onthese (Eid et al. 2004; González-Ferreiro et al. 2012; Wehr andLohr 1999). LiDAR is an active remote sensing system basedon the use of a laser sensor and the application of various tech-niques to determine the distance from a laser transmitter to anobject Sánchez Martínez et al. (2011). This distance isestablished by measuring the time delay between emission ofa signal and detection of its reflection (Tanarro 2010). Themeth-od is therefore a combination of three different technologies:laser telemetry, the global positioning system (GPS), and inertialmeasurement units (IMUs). The laser beam, emitted at a fre-quency of thousands of energy pulses per second toward theearth, creates a dense strip of 3D points (Manue 2007). This3D point cloud is based on accurate measurement by a plane-mounted pulse sensor, which calculates the distance separatingit from the earth’s surface and objects existing on it (Magdaleno-Mas and Martínez-Romero 2006). As the position and orienta-tion of the sensor are known for each pulse emitted, each returnsignal has unique three-dimensional coordinates. LiDAR datahave been captured for the entire Spanish territory under theNational Aerial Orthophotography Plan (PNOA). Data werecollected during 2012 in the region of Asturias.

LiDAR technology has been used successfully to charac-terize numerous types of forest stands (Hayashi et al. 2014;Lefsky et al. 2002;Means et al. 1999;Means andAcker 2000;Næsset et al. 2004). However, in forest inventories, the tech-nique has been found to underestimate height (Clark et al.2004; Næsset 1997; Zimble et al. 2003), and some authors(Falkowski et al. 2006)have suggested thathigherdensitydata(6–8 pulsesm−2) are required for forestmonitoring. However,other studies have shown that a low pulse density is sufficientfor establishing strong correlations with the main attributesmeasured in forest inventories (Hawbaker et al. 2010; Meansand Acker 2000; Thomas et al. 2006). Although studiesbased on the use of low-density discrete-return LiDARto determine forest structure have been reported (Coopset al. 2007; Hall et al. 2005), to date only basal areahas been estimated in short-rotation coppice (SRC)(Seidel and Ammer 2014). Nonetheless, the structureof SRCs facilitates the application of LiDAR technologyas these are dense, rather uniform stands with little orno accompanying vegetation. These features favor goodcorrelations between variables measured in forest inven-tories and those measured using LiDAR technology.

The main objective of this study was to assess the useful-ness of low-resolution discrete return LiDAR (0.5 pulses m−2)data to estimate structural attributes and biomass productionwith the aim of facilitating management of an SRC plantation.For the purposes of this study, we developed statistical modelsthat relate the information provided by the LiDAR to the dataobtained in detailed field inventories conducted in the studyarea (Montealegre Gracia et al. 2015). The methods wereevaluated by complementary techniques: parametric multiplelinear regression, which enabled us to develop predictivemodels, evaluated by Rfit

2 and RMSE in order to indicatethe accuracy of the fits, and non-parametric classificationand regression trees (CART), which provided more detailed,descriptive information about the variables. The main aims ofthe study were (i) to estimate the forest structure and the pro-ductivity of SRWC and (ii) to apply and compare the use ofdifferent types of model fitting methods (multiple linear re-gression and CART).

2 Material and methods

2.1 Study area

2.1.1 Location of the study area

The experimental trial included three commercial willowclones and covered an area of ≃ 2 ha in the region ofAsturias (north-western Spain) (Fig. 1). The Salix energy croptrial was established inMay 2008 in restored land surroundingan abandoned opencast coal mine, denominated Mozquita(ETRS89 UTM 30 N, N: 4,794,443, E: 280,981). The studyarea is characterized by an average annual temperature of13 °C and an average annual precipitation of 1,115 mm, ofwhich 345 mm falls during the growing season (May–September). The climate is oceanic with high annual precipi-tation and, although summer precipitation is relatively low insome areas, physiological drought does not occur in any partof the region, which is located entirely within the EuropeanBiogeographic Atlantic Region (EEA 2011).

The clay loam substrate (with a high presence of coarseelements, approximately 30%) was dumped and amelioratedin 2003. Soil formation is at an early stage and the soil struc-ture is still unstable. The steep slopes of the terrain minimizegroundwater effects. The physiography of the plots was char-acterized by a mean slope of 19% and an elevation rangingfrom 508 to 597 m above sea level.

2.1.2 Experimental design

In the winter of 2008, the surface was subsoiled, plowed to adepth of 30–40 cm, and harrowed before the willow cuttingswere planted. Three commercially available willow clones

Annals of Forest Science (2017) 74: 69 Page 3 of 16 69

were chosen for the study because of their adaptability toextreme soil conditions (e.g., nutrient poor and polluted soils)(Kuzovkina and Quigley 2005) and because they display goodstructural attributes and yield capacities for biomass produc-tion in SRC (Keoleian and Volk 2005). The cuttings wereplanted according to a double row planting design, leaving adistance of 0.75m between each set of double rows, a distanceof 1.5 m to the next set of double rows, and a distance betweenplants of 0.9 m (10,000 plants ha−1) or 0.6 m (15,000 plantsha−1) to provide two stocking levels (Fig. 2).

The experiment was established following a randomizedcomplete block design (three blocks), in which three qualita-tive factors were considered for analysis: clone (three levels),planting density (two levels), and fertilization treatment (threelevels), as outlined in Table 1.

Finally, the basic design was repeated in three blocks, witha total of 54 square plots each with an area of 400 m2

(20 × 20 m) and constituted by 9 double rows with 22 or 33cuttings per row (depending on the stocking density).Irrigation and pest/disease control were not performed duringthe cultivation period throughout the study area.

2.2 Field data collection

The experimental plots were measured in the autumn of2012 according to the protocol described by the UKForest Research (Forest Research 2003) for collecting datain short-rotation willow plantations (first rotation, standage 5 years). Several variables were measured after the

vegetative period with the aim of assessing the perfor-mance of each clone. Measurements were made in rectan-gular subplots of 27 m2 (9 × 3 m for density N1) and 18 m2

(6 × 3 m for density N2) located in the center of each plot,to avoid the edge effect. A total of 40 stools (live or dead)were measured in each of the 54 subplots in this study.Within each of the subplots, shoot diameters were mea-sured 20 and 130 cm aboveground level (D20 and D130)with a digital caliper, and the total mean heights (H) weremeasured with a Vertex III hypsometer. The survival ratewas also recorded at the end of each vegetative season.Before the trial, crops were harvested (in autumn 2012;stand age 5 years) and 5 stools were randomly selected ineach of the abovementioned subplots and subsequently cut.

Fig. 1 Distribution of thesampling plots used forestimation of structural attributesand biomass yield. Thephotographic insert shows theexperimental layout of the threecommercial willow clones understudy (green, Bjor clone; red,Inger clone; and blue, Olof clone)

Fig. 2 Diagram of the planting designs used in the trial

69 Page 4 of 16 Annals of Forest Science (2017) 74: 69

A total of 270 stools were harvested manually with pruningshears (Electrocoup F3010) or a chainsaw. The freshweight (FW) of each stool was measured with an electronicbalance (precision ± 10 g). A representative subsample(300–500 g) of each stool was taken to the laboratory andweighed immediately with a precision balance (precision± 1 g). The subsample was subsequently dried to constantweight at 70 °C: the dry weight (DW) was recorded and thedry biomass of the plot was calculated by multiplying theFW by the ratio of dry to fresh weight of the subsample.

The mean values and standard deviations for thesevariables are shown in Table 2. The biomass and struc-tural attributes differed depending on the treatment ap-plied (F0, F1, and F2) (Table 1), thus explaining thehigh standard deviations. The trial involved a denseplantation (10,000 and 15,000 plants ha−1) with a veryhomogeneous stand structure in each plot and scarcelyany companion vegetation. The plot location data (fourvertices) were collected with GPS submeter precision(Trimble Geoexplorer 2008 series).

2.3 LiDAR procedure

2.3.1 LiDAR data

The LiDAR data (Table 3) were acquired in July 2012 withALS60 (Leica) and LMS-Q680 (Riegl) sensors. The beamdivergence was 0.3 mrad and the pulsing frequency, 45 kHz;the scan frequency was 70 Hz, and the maximum scan angle,50°. The first and last return pulses were recorded. Flightswere conducted across the whole study area, and one flightwas conducted for each strip, yielding an average measure-ment density of about 0.5 pulses m−2. The LiDAR data pro-vided by the PNOA (official web page, 2016) included infor-mation about return type (first and last); X, Y, and Z coordi-nates; and intensity of the returned pulse by the sensor.

2.3.2 Extraction of LiDAR variables

For the low-density data acquired (0.5 pulses m−2), FUSIONsoftware (McGaughey 2010) was used to filter, interpolate,

Table 1 Main characteristics ofthe experimental design Characteristic Trial

Total area 2.3 ha

Experimental design Randomized complete blocks

Number of replicates 3

Species Salix spp.

Clones tested Bjor (B), Inger (I) and Olof (O)

Origin of clones Sweden

Progenitor B: Salix schwerinii × Salix viminalis

I: Salix trianta × Salix viminalis

O: Salix viminalis × (Salix schwerinii × Salix viminalis)

Plant density/spacingbetween cuttings

N1 = 10,000/0.9 N2 = 15,000/0.6 (plants·ha−1/m)

Treatment F0 F1 F2

Fertilization None 300 kg ha−1 N-P-K 6:20:12 600 kg ha−1 N-P-K 6:20:12

Herbicide None Application of glyphosate (4 l/ha) Application of glyphosate (4 l/ha)

Table 2 Mean values andstandard deviations of the testvariables (H, D20, D130, FW, andDW) for the three clones (datacorresponding to a field inventorycarried out in 2012)

Clone Treatment H (m) D20 (cm) D130 (cm) FW (t ha−1) DW (t ha−1)

Bjor F0

F1

F2

1.4 ± 0.4

2.4 ± 1.2

2.9 ± 0.5

1.1 ± 0.3

1.8 ± 0.9

2.2 ± 0.5

0.6 ± 0.2

1.2 ± 0.5

1.4 ± 0.4

1.8 ± 0.6

8.1 ± 6.6

10.9 ± 3.5

0.76 ± 0.4

4.45 ± 3.7

5.88 ± 2.1

Inger F0

F1

F2

1.7 ± 0.7

2.7 ± 0.8

2.7 ± 0.7

1.4 ± 0.6

1.9 ± 0.3

2.0 ± 0.4

1.0 ± 0.5

1.2 ± 0.3

1.3 ± 0.3

4.9 ± 3.7

17.1 ± 7.7

14.7 ± 8.3

2.54 ± 1.9

9.32 ± 4.4

8.23 ± 5.0

Olof F0

F1

F2

2.7 ± 0.7

5.7 ± 0.5

5.5 ± 1.4

1.7 ± 0.3

3.2 ± 0.3

3.0 ± 0.7

1.3 ± 0.2

2.4 ± 0.2

2.3 ± 0.6

6.4 ± 4.5

34.8 ± 11.9

33.2 ± 10.8

3.60 ± 2.6

19.71 ± 6.7

18.88 ± 6.9

Annals of Forest Science (2017) 74: 69 Page 5 of 16 69

and generate the digital terrain model (DTM)/digital crownmodel (DCM) and also to compute the following variablesrelated to the metrics of heights and return intensity distribu-tions within the limits of the 54 field plots: mean, maximumand minimum values, mode, standard deviation, variance, in-terquartile distance, coefficients of skewness and kurtosis, av-erage absolute deviation, and percentiles. The proportion ofreturns (%) above a specific height threshold was also estimat-ed in order to separate tree crop canopy returns from othervegetation.

The following steps were carried out with several process-ing programs (algorithms) implemented in the Fusion LiDARToolkit (McGaughey 2010).

– GROUNDFILTER. Ground returns were extracted fromthe LiDAR point cloud by using the GroundFilter tool,which implements a filtering algorithm adapted fromKraus and Pfeifer (1998) based on linear prediction(Kraus and Mikhail 1972).

– GRIDSURFACE CREATE. These returns were used togenerate a DTM (1 m2 resolution) grid with theGridSurfaceCreate tool, which computes the elevationof each grid cell by using the average elevation of allpoints within the cell; if the cell does not contain groundreturn points, its elevation is generated by interpolationfrom the neighboring cells. The metrics were generatedfor the exact size of each 20 × 20 m plot.

– CLIPDATA. The normalized LiDAR point cloud was ob-tained by subtraction of the ellipsoidal height of the DTMfrom the Z coordinate of each LiDAR return with theClipData tool; this tool was used also to exclude returnsbelow a normalized height of 0.5 m, considered as not be-longing to tree crowns (e.g., hits on shrubs, rocks and logs).

– POLYCLIPDATA. ThenormalizedLiDARpointcloudwasclippedwith the limitsofeach fieldplot—whichwerestoredas polygons in vector format—by using the PolyClipDatatool; one independent file was created per plot.

– CLOUDMETRICS. The metrics of heights and return in-tensity distributions of the 54 clipped and normalizedpoint clouds were computed with the CloudMetrics tool.

Table 4 shows the complete set of metrics and the corre-sponding abbreviations used in this study.

2.4 Statistical analysis

Several statistical methods can be used in remote sensingprediction studies. In this study, parametric and non-parametric fitting methods were used to predict the mainstructural attributes and biomass production variablesfrom LiDAR data and field measurements. Both fittingmethods were used for comprehensive analysis of the dataand because they provide important complementaryinformation.

Table 3 Specifications of the LiDAR flight. Source: PNOA

Parameter Specification

Pixel dimension 0.25 m

Point density 0.5 pulses m−2

Geodetic reference system ETRS89

Map projection UTM

Pulse rate 45 kHz

Areas covered 1:5000

Sensors ALS60 (Leica); LMS-Q680 (Riegl)

Table 4 Statistics extracted from the heights and intensities of LiDARflights and used as independent variables for the regression models

Description Abbreviation

All returns above 1.00 All returns above 1.0

All returns above mode All returns above mode

Canopy relief ratio Canopy relief ratio

Elevation kurtosis Elev kurt

Elevation L1 Elev L1

Elevation minimun Elev minimun

Elevation mean Elev mean

Elevation stddev Elev stddev

Elevation 30th percentile Elev P30

Elevation 40th percentile Elev P40

Elevation 50th percentile Elev P50

Elevation 60th percentile Elev P60

Elevation 70th percentile Elev P70

Elevation 75th percentile Elev P75

Elevation skewness Elev skewness

Elevation variance Elev variance

Int sttdev Int sttdev

Intensity kurtosis Int kurt

Intensity L4 Int L4

Intensity of 5th percentile Int P05

Intensity of 10th percentile Int P10

Intensity of 25th percentile Int P25

Intensity of 30th percentile Int P30

First return above mode First return above mode

Percentage all returns above 1.00 % all returns above 1.0

Percentage all returns above mean % all returns above mean

Percentage first returns above 1.00 % first returns above 1.0

Return 1 count above 0.50 Return 1 count above 0.50

Total return count Total return count

Total first returns Total first returns

Total returns count above 0.50 Total returns count above 0.50

69 Page 6 of 16 Annals of Forest Science (2017) 74: 69

2.4.1 Parametric methods

Multiple linear regression (MLR) was used to model the rela-tionships between field measurements (H, D20, D130, FW, andDW) and the LiDAR variables in order to produce generalmodels (for all clones together) and models classified by clone(Bjor, Inger, and Olof).

Candidate predictor variables were required to have an en-tering F-statistic with a significance level of 0.05 or less forinclusion in the model, and no predictor was left in the modelwith a partial F-statistic with a significance level greater than0.05. Dependent variables derived from field data and predict-ed in regressions were mean height (H), basal diameter (D20),diameter at breast height (D130), fresh weight (FW), and dryweight (DW).

Comparison of the model estimates was based on the fol-lowing two statistics: the adjusted coefficient of determination(Rfit

2) and the root mean square error (RMSE). The Rfit2 com-

pares the descriptive power of regression models that includethe diverse numbers of predictors, and RMSE is a quadraticscoring rule that measures the average magnitude of the error(the square root of the average of squared differences betweenpredicted and actual observations), which was calculated toprovide additional information. Finally, residual plots werechecked in order to validate the model fit. The variance infla-tion factor (VIF) was also taken into account. This factorquantified the severity of multicollinearity in ordinary leastsquares regression analysis and also provided an index thatmeasured the extent to which the variation in an estimatedregression coefficient increased due to collinearity. Onlymodels in which all parameters were significant at the 5%level and with a VIF < 5 were included, thus ensuring thatpredictions were not highly correlated (Belsley et al. 2005;Mandeville 2008).

In this study, all of the data were used to construct thegeneral and clone-specific models, as according to Myers(1990, p. 170) and Hirsch (1991), final estimation of modelparameters using the entire dataset is more precise thanwhen amodel is fitted using only one portion of the data, especiallywith relatively small sample numbers.

2.4.2 Non-parametric method

In a preliminary analysis carried out to determine the mostappropriate non-parametric statistical method, random forest(RF) and classification and regression trees (CART) werecompared. The CARTmethodwas chosen because it providedbetter fits to the data, with larger R2 values and lower RMSE,than yielded by RF (see Online Resource 4). It is also a goodexploratory technique that aims to determine classificationand prediction rules.

The main advantages of the CART method are as follows(Gordon 2013; Timofeev 2004): (i) it does not require

specification of any functional form, (ii) it does not requirevariables to be selected in advance, (iii) it can easily handleoutliers, (iv) it does not require the assumptions of statisticalmodels and is computationally fast, (v) it is flexible and candeal with missing data, (vi) it works better than RF with rela-tively small databases, and (vii) it is easy to interpret (unlikerandom forest).

The objective of CART is usually to classify a dataset intoseveral groups by use of a rule that displays the groups in theform of a binary tree (Breiman et al. 1984), which is deter-mined by a procedure known as recursive partitioning. In thisstudy, the CART method was used to classify the variablesconsidered (H, D20, D130, FW, and DW) in relation to theLiDAR data available.

Each tree branch is described by the value of one descrip-tor, chosen so that all objects in a daughter group have moresimilar response variable values. The split for continuous var-iables is defined by xi < aj, where xi is the selected descriptoror explanatory variable, and aj is its split value. To choose themost appropriate descriptor xi and value of aj, CART uses analgorithm in which all descriptors and all split values are con-sidered, selecting those giving the best reduction in impuritybetween the mother group (tp) and the daughter groups (tL andtR) (Deconinck et al. 2005). This process is repeated for eachdaughter group until the maximal tree height is reached.Mathematically, this is expressed as follows:

Δi s; tp� � ¼ i tp

� �−pLi tLð Þ−pRi tRð Þ ð1Þ

where it is the impurity, s the candidate split value, and pL andpR are the fractions of the objects in respectively the left andright daughter groups.

The impurity is defined as the total sum of squares of thedeviations of the individual responses from the mean responseof the group and is expressed as follows:

i tð Þ ¼ ∑n

yn−y tð Þ� �2

ð2Þ

where i(t) is the impurity of group t, yn is the value of theresponse variable for object xn, and y tð Þ is the mean of theresponse variable in group t.

CART methods are not required to conform to probabilitydistribution restrictions, and there is no assumption of linearityor any need to pre-specify a probability distribution for theerrors (Bell 1999).

Complexity and robustness are competing characteristicsthat must be considered simultaneously during constructionof statistical models. The more complex a model is, the lessreliable it will be for purposes of prediction. To prevent thisfrom occurring, stopping rules must be applied during elabo-ration and the development of a decision tree to prevent themodel from becoming overly complex. Common parameters

Annals of Forest Science (2017) 74: 69 Page 7 of 16 69

used in stopping rules include (a) the minimum number ofobservations in a leaf, (b) the minimum number of observa-tions in a node prior to splitting, and (c) the depth (i.e., numberof levels) of any leaf from the root node (Song and Lu 2015).On this occasion, no pruning was necessary because a maxi-mum of 2 levels was considered for tree depth (or a maximumof 6 nodes). The risk estimate, which is a measure of thewithin-node variance, was used as an indicator of model per-formance (IBM Corp. Released 2015).

3 Results

We used two well-defined and complementary parametric andnon-parametric fitting procedures to analyze and estimate thebest response. In this case, the structural attributes (H, D20,and D130) and biomass yield variables (FW and DW) werestrongly and positively correlated (R2 > 0.87, additional dataare given in Online Resource 5).

The main results obtained with both fitting methods areshown in Tables 5, 6, and 7. All the dependent variables (H,D20, D130, FW, and DW) were analyzed and interpreted at triallevel and separately for each clone (see Table 7 for data on theOlof clone and additional data are given in Online Resources 1and 2 for data on the Bjor and Inger clones). The parameterestimates and goodness-of-fit statistics for the best parametricmodel (multiple linear regression) are summarized in Table 5.Finally, the best non-parametric (CART) models are includedin Table 6 (trial level) and Table 7 (Olof clone). The scatterplots generated byMLR and CART (Online Resource 3) showthe generally close relationship between the predicted valuesand the field measured values.

3.1 Parametric methods

At trial level, the models for the structural attributes (H, D20,and D130) provided good fits, with more than 76% of thevariance explained, and the LiDAR variables with greatestinfluence were those related to elevation percentile, totalreturns, and the first returns on 0.5. The highest R2 valuewas obtained for the mean height variable (Rfit

2 = 89%,Table 5). When modeling the biomass variables (FW andDW), the Rfit

2 values were higher than 75% (Table 5), andin both cases, the most influential variables in the model wereelevation percentile and percentage of all returns above mean.

Regarding the fit for the models for each clone, the Bjorclone model provided the best results with H and D20 (91 and97%, respectively; Table 5). However, the Olof clone modelproduced the highest Rfit

2 using D130 (95%; Table 5). In thecase of the biomass variables (FW and DW), the amount ofvariance explained was sometimes lower, with Rfit

2 valuesabove 66%. The Bjor clone model produced the highest Rfit

2

values (90%) for both the FW and DW variables (Table 5).

The homogeneity of variance was evaluated using SPSSgraphs obtained with ZRESID and ZPRED (standardized pre-dicted values and standardized residuals) commands. Plots ofresiduals against predicted values showed no evidence of het-erogeneous variance and no systematic pattern. The resultsindicated the absence of atypical or scattered sample dataand, furthermore, bell-shaped histograms indicated that alldatasets were normally distributed.

3.2 Non-parametric methods

The proportion of variance explained by the mean heightvariable-dependent model for the whole trial produced an R2

value of 84.6%. In this case, the most important LiDAR var-iable was the mean elevation. It is important to highlight theinfluence of this variable on the other structural attributes andproduction variables, apart fromD20, with the most significantindependent variable related to the elevation of the percentiles,namely the variable Elev. P60 (Table 6).

The independent variables that best define the decisiontrees (CART) per clone weremore varied. The fitting providedgood results for all three clones studied. For the Olof clone,which yielded the highest values for structural attributes andbiomass production in the field, the fit was good, withR2 > 92% for all study variables (Table 7). However, in plotswith lower values for mean height, diameter, and biomass, i.e.,plots with the Bjor and Inger clones, the independent explan-atory variables were those related to different percentiles ofelevation and % returns (Online Resources 1 and 2). In theCART analysis for each clone, we observed that, with someexceptions, most of the models produced a second level ofclassification, with variables related to elevation and %returns.

Once the fitting was completed, we were able to verify thatthe LiDAR variables associated with mean elevation and ele-vation percentile (tree height) were the most important in thefitting process, as these were included in all models producedby the parametric and non-parametric fitting methods used.

4 Discussion

The parametric and non-parametric model fitting conducted inthis study has revealed acceptable results that indicate theusefulness of LiDAR data for estimating structural attributesand biomass production in forest energy crops grown on de-graded mining land.

The data collected in the field inventory and LiDAR data(available for free) captured in the same period were highlycorrelated. One of the objectives of the study was to determinedifferences in the performance of the different models studied(H, D20, D130, FW, and DW) at trial level (without differenti-ating clones) and separately for each clone.

69 Page 8 of 16 Annals of Forest Science (2017) 74: 69

Table 5 Results of multiple regression showing the best models obtained for H, D20, D130, FW, and DW, for each trial and each clone

Dependentvariable

Factor Modeltype

Independent variable Parameterestimate

Std.Error

RMSE R2 Rfit2

(%)

H (m) Trial Multiple (Constant)Elev P50Total return countReturn 1 count above 0.50

3.4960.810− 0.0090.006

0.7710.0760.0020.002

0.54 0.90 89.00

Bjor clone Multiple (Constant)% all returns above 1.00Return 1 count above 0.50Elev skewnessElev variance

1.1160.324− 0.059− 1.1590.712

0.1760.0350.0090.2072.222

0.29 0.93 91.00a

Inger clone Multiple (Constant)% first returns above 1.00

1.3600.041

0.2100.07

0.51 0.70 68.00

Olof clone Multiple (Constant)Elev P50Int P05

− 9.1611.2340.900

3.8470.1230.350

0.60 0.88 87.00

D20 (cm) Trial Multiple (Constant)Elev P50Total return countReturn 1 count above 0.50

2.2620.336− 0.0040.003

0.5670.0560.0010.001

0.40 0.76 75.00

Bjor clone Multiple (Constant)ElevP25Total return countElev P01Canopy relief ratioFirst returns above 1.00Lnt mode

4.5690.574− 0.011− 1.6173.1110.010−0.046

0.3430.0970.0010.2660.4390.0020.019

0.12 0.98 97.00a

Inger clone Multiple (Constant)Elev P60

0.2470.856

0.2900.158

0.30 0.65 63.00

Olof clone Multiple (Constant)Elev P50Int P10Int P30Int P25

− 3.4360,5460.501−0.5700.467

1.2530.0450.1720.1760.179

0.24 0.93 91.00

D130 (cm) Trial Multiple (Constant)Elev P30Return 1 count above 0.50Total return count

1.3150.3750.002− 0.002

0.4180.0470.0010.001

0.29 0.82 81.00

Bjor clone Multiple (Constant)Total first returnsElev P50Canopy relief ratio

0.521− 0.0061.0071.191

0.1680.0010.1260.540

0.17 0.91 89.00

Inger clone Multiple (Constant)Elev meanTotal first returnsInt skewness

2.0520.449− 0.0060.215

0.5610.1310.0010.094

0.19 0.81 77.00

Olof clone Multiple (Constant)Elev P50Int P10Int P30Elev P40Int P10

− 33.6011.4180.579− 0.258− 1.0273.077

12.7880.3290.1040.0690.3511.319

0.14 0.96 95.00a

FW (t ha−1) Trial Multiple (Constant)Elev P60%all returns above mean

− 4.8944.6820.596

1.9540.8590.122

6.64 0.76 75.00

Bjor clone Multiple (Constant)% all returns above 1.0First returns above 1.0Elev skewnessElev L4

2.7772.326− 0.522− 3.9929.128

0.8190.3660.1110.9203.725

1.84 0.92 90.00a

Inger clone Multiple (Constant)(All returns above mode) / (Total first returns) × 100Int kurt

− 1.3550.4891.092

2.1700.0660.374

4.03 0.79 77.00

Annals of Forest Science (2017) 74: 69 Page 9 of 16 69

The results indicate that free LiDAR data can be used toestimate the variables with acceptable precision (Rfit

2 > 63%inmultiple regression and R2 > 73.5% for CARTanalysis) andthat satisfactory results were obtained, despite the low densityof LiDAR points. At the trial level, although MLR producedhigher R2 values than CART, except for the D20 and FW fits,the values were generally very similar. However, when the fitswere carried out taking into account the clone level (i.e., amore uniform sample), the CART method produced largerR2 values (except for D130 in Inger and Olof and H and D20

inBjor). Finally, CARTandMLR producedverysimilarRMSEvalues at trial level, but CART generally yielded smaller RSMEvalues at the clone level (except forBjor andOlof cloneswithD20

andD130, respectively). Thegoodness-of-fit levels yielded by themodels are comparable to those obtained in other studies usingLiDAR to characterize forest attributes (Dalponte et al. 2011;García-Gutierrez et al. 2014; Seidel andAmmer 2014).

The MLR and CART scatter plots (additional data aregiven in Online Resource 3) show that both methods providedacceptably good fits to the data. However, for the structuralvariables (H, D20, D130), the models developed produced bet-ter predictions than for biomass variables (FW and DW), inthat the data were more widely dispersed.

On the other hand, to verify and compare the models, forboth parametric and non-parametric methods, a newly collect-ed dataset (additional inventory) should be used for validation(Hirsch 1991; Kozak and Kozak 2003; Myers 1990) becausethe only universally acceptable method for validating a modeland assessing its goodness of fit after model selection is to usean independent sample (Lever et al. 2016). However, indepen-dent validation was not possible in this study, as application ofa validation method requires more data and the CARTmethod

thus becomes more unstable (Gordon 2013), especially asregards the models developed for each clone. Nonetheless,the models fitted in this study can be considered sufficientlyrobust for estimating structural attributes and productivity ofSRWC. In this case, the CART analysis of the relationshipsbetween field-measured variables and LiDAR data producedreasonable results. Moreover, the errors were acceptable, con-sidering the high degree of variability between the trial plots.

The trial level models produced by the MLR method showthat the most important variables, for trial and clone, werethose related to elevation and returns, as also shown by theCART method. However, the models generated by MLR foreach clone were defined by more diverse LiDAR variables,which also included (apart from those already mentioned)variables related to intensity (see Table 5).

Good fits were obtained for the structural attributes studied,mean height (H) and both diameters (D20 and D130), probablybecause these are closely related to the LiDAR-derived eleva-tion variables (tree height). However, good-fit models werealso obtained for FWand particularly DW biomass, as expect-ed, because these variables are strongly correlated with struc-tural variables (see Online Resource 5). Some variables suchas mean elevation (Elev mean) and elevation percentile (Elev%) were included in most of the models. The data for the trialplots planted with the Olof clone were uniformly distributed,and the independent variable that provided the best predic-tions was the mean elevation (Elev mean).

The mean height variable was closely correlated with theLiDAR variables associated with the height of the trees (i.e.,elevation) and with the returns, as also observed byMontealegreGracia et al. (2015) in a study of aPinus halapensisMill plantation in Spain. Another study in plantations of

Table 5 (continued)

Dependentvariable

Factor Modeltype

Independent variable Parameterestimate

Std.Error

RMSE R2 Rfit2

(%)

Olof clone Multiple (Constant)Elev P75Int P90Elev P70Int L4

− 67.239100.2161.827− 89.366− 22.947

12.76025.3100.49625.0768.855

6.35 0.88 85.00

DW (t ha−1) Trial Multiple (Constant)Elev. P50%all returns above mean

− 2.9212.8170.355

1.0430.4930.067

3.67 0.78 77.00

Bjor clone Multiple (Constant)% all returns above 1.0First returns above 1.0Elev skewnessElev kurt

0.2101.248− 0.289− 2.6150.751

0.6150.2170.0640.5970.296

1.05 0.92 90.00a

Inger clone Multiple (Constant)(All returns above mode) / (Total first returns) × 100Int kurt

− 1.1840.2830.633

1.1900.0360.205

2.21 0.81 79.00

Olof clone Multiple (Constant)Elev P75

− 8.0934.917

3.9750.835

5.42 0.68 66.00

a Best fit at the trial level and per clone for each study variable. The RMSE units are the same as the variable units

69 Page 10 of 16 Annals of Forest Science (2017) 74: 69

Tab

le6

Resultsof

thenon-parametricfitting

(CART)of

themodelsobtained

forH,D

20,D

130,F

W,and

DW,atthe

triallevel

Variable

R2

(%)

RMSE

Node

Mean

Standard

deviation

Num

ber

Percent

Predicted

mean

Parental

Node

Independentv

ariable

Variable

Improvem

ent

Value

H(m

)84.6a

0.64

03.14

1.65

53100

3.14

––

––

12.41

0.90

4279.2

2.41

0Elevmean

1.98

≤3.64

25.89

0.62

1120.8

5.89

>3.64

31.65

0.65

1630.2

1.65

1%

firstreturnsabove1.00

0.28

≤10.09

42.88

0.69

2649.1

2.88

>10.09

D20(cm)

76.4

0.38

02.04

0.80

53100

2.04

––

––

11.63

0.49

3769.8

1.63

0ElevP60

0.38

≤2.58

22.98

0.52

1630.2

2.98

>2.58

31.04

0.29

1018.9

1.04

1%

firstreturnsabove1.00

0.09

≤5.83

41.85

0.34

2750.9

1.8

>5.83

D130(cm)

79.1

0.30

01.44

0.66

53100

1.44

––

––

11.16

0.42

4279.2

1.16

0Elevmean

0.28

≤3.64

22.48

0.25

1120.8

2.48

>3.64

30.56

0.20

713.2

0.56

1Totalreturns

above0.50

0.06

≤15.00

41.28

0.34

3566.0

1.28

>15.00

FW

(tha

−1)

78.8

6.47

014.90

13.22

53100

14.94

––

––

19.52

7.32

4279.2

9.52

0Elevmean

110.56

≤3.64

235.45

10.17

1120.8

35.45

>3.64

30.56

0.20

713.2

0.56

1Firstreturnsabove1.00

19.10

≤38.00

41.28

0.34

3566.0

1.28

>38.00

DW

(tha

−1)

77.2

3.64

08.15

7.63

54100

8.15

––

––

15.05

4.14

4379.6

5.05

0Elevmean

37.54

≤3.64

220.26

5.86

1120.4

20.26

>3.64

33.69

2.89

3564.8

3.69

1%

firstreturnsabove1.00

6.43

≤25.80

411.00

3.53

814.8

11.0

>25.80

aBestfitatthetriallevelforeach

studyvariable.T

heRMSE

units

arethesameas

thevariableunits

Annals of Forest Science (2017) 74: 69 Page 11 of 16 69

Tab

le7

Resultsof

non-parametricfitting

ofthemodelsobtained

forH,D

20,D

130,F

W,and

DW,for

theOlofclone

Variable

R2

(%)

RMSE

Node

Mean

Standard

deviation

Num

ber

Percent

Predicted

mean

Parental

Node

Independentv

ariable

Variable

Improvem

ent

Value

H(m

)94.2

0.40

04.68

1.68

18100.0

4.68

––

––

12.77

0.69

738.9

2.77

0Elevmean

2.30

≤3.64

25.89

0.62

1161.1

5.89

>3.64

36.5

0.34

316.7

6.57

2Intsttdev

0.10

≤2.52

45.64

0.51

844.4

5.64

>2.52

D20(cm)

93.6

0.20

02.62

0.81

18100

2.62

––

––

11.70

0.29

738.9

1.70

0Elevmean

0.53

≤3.64

23.20

0.31

1161.1

3.20

>3.64

31.50

0.17

422.2

1.50

1%

firstabovemode

0.02

≤12.83

41.96

0.20

316.7

1.96

>12.83

53.61

0.24

211.1

3.61

2%

firstabove1

0.02

≤32.11

63.11

0.25

950.0

3.11

>32.11

D130(cm)

94.8a

0.15

02.01

0.65

18100.0

2.01

––

––

11.26

0.19

738.9

1.26

0Elevmean

0.35

≤3.64

22.48

0.25

1161.1

2.48

>3.64

31.09

0.08

316.7

1.09

1Totalreturncount

0.00

≤47.50

41.40

0.13

422.2

1.40

>47.50

52.85

0.19

211.1

2.85

2%

firstreturnsabove1.0

0.01

≤32.11

62.40

0.19

950.0

2.40

>32.11

FW(tha

−1)

92.3

4.48

024.81

16.20

18100.0

24.81

––

––

18.09

6.12

738.9

9.09

0%

allreturnabovemean

177.89

≤3.64

235.45

10.17

1161.1

35.45

>3.64

34.77

2.46

527.8

4.44

1%

firstreturn

abovemode

10.74

≤49.00

416.40

2.76

211.1

16.40

>49.00

552.43

4.01

211.1

51.43

2Elevminimun

39.16

≤39.16

631.68

6.28

950.0

31.69

>39.16

DW

(tha

−1)

92.2

2.60

014.06

9.36

18100

14.06

––

––

14.31

3.07

738.9

4.31

0Elevmean

60.49

≤3.64

220.2

5.86

1161.1

20.26

>3.64

32.65

1.48

527.8

2.65

1Firstreturn

abovemode

2.65

≤49.00

48.44

0.22

211.1

8.44

>49.00

529.99

1.19

211.1

29.99

2Elevminimun

12.85

≤0.53

618.10

3.37

950.0

18.10

>0.53

aBestfitperOlofclone

foreach

studyvariable.T

heRMSE

units

arethesameas

thevariableunits

69 Page 12 of 16 Annals of Forest Science (2017) 74: 69

Pinus radiata D.Don in northern Spain (González-Ferreiro et al.2012) indicated that mean height can be accurately modeledfrom low-density laser data (0.5 pulses m−2), yielding an R2

value of 0.75 (in comparison with R2 values of 0.70–0.93 ob-tained by parametric methods and R2 values of 0.85–0.94 bynon-parametric (CART) methods in the present study). In astudy of an olive plantation in Spain, with low-density data(0.5 pulses m−2), good results were also obtained for height(R2 = 0.67) (Estornell et al. 2014).

The results obtained for the models of the diameter variables(D20 and D130) were better than expected, taking into accountthe reports in the relevant scientific literature. Thus, the modelsused to estimate diameters performed similarly to those reportedby Gonçalves-Seco et al. (2011) for Eucalyptus globulus Labillplantations in Spain (R2 = 0.71) and byDalponte et al. (2011) fora mixed stand (R2 = 0.63). Similar findings were also reportedby Graham (2008) for the D130 model and a pine plantation(R2 = 0.82). However, in the same study, a much lower R2 valuewas obtained for a natural pine stand (R2 = 0.20). In a studyconducted by Estornell et al. (2014) in a plantation of small trees(olives) in the Mediterranean area of Spain, with low-densityLiDARdata (0.5 pulsesm−2), the findings indicated goodmodelperformance (R2 = 0.70) for estimating volume, which is direct-ly related to diameter. The findings thus seem similar to those ofthe present study.

Several studies carried out worldwide with low-point-density data have reported similar results to those obtainedin the present study. For example, a study carried out inEurope to evaluate the aboveground biomass in boreal forestzone with an average point density of between 0.7 and 1.2 m−

2 reported an R2 value of 0.88 (Næsset and Gobakken 2008).A study carried out in Canada using low-density data (0.5pulses m−2) obtained an excellent fit for the biomass, withan overall R2 of 0.93 (Treitz et al. 2010). Li et al. (2008) alsoobserved a significant relationship between field-based above-ground biomass estimates based on field and LiDARmeasurements for the three study sites, located in the USAand Canada and in which different forest species were used.However, a study by Næsset (2011) in small areas of forestland in Norway where the main tree species were Norwayspruce (P. abies (L.) Karst.) and Scots pine (P. sylvestris L.)showed that these species are subject to substantial inherentcanopy height variation, leading to highly variable predictionsfor estimate aboveground biomass in young forests(Magnussen and Boudewyn 1998).

In the present study, no comparison was made for differentLiDAR point densities because of the scarcity of the LiDARdata. Previous studies did not find any evidence indicating thata reduction in point density affects the model accuracy, andTreitz et al. (2010) considered that data captured at 0.5 pulsesm−2 may be an excellent source of information for forest man-agement. The same was also concluded by González-Ferreiroet al. (2013), who showed that low-intensity LiDAR data (0.5

pulses m−2) can be used without significant loss of informa-tion. This suggests that data captured at 0.5 pulses m−2 yieldsgood estimates, without excessive loss of model quality.

According to the findings reported by García et al. (2010),the intensity variables are more strongly related to biomassthan to mean height, as also shown in the present study.Likewise, González-Ferreiro et al. (2013) found that the inde-pendent variables associated with the return intensity and re-lated canopy measurements can add some valuable informa-tion for predicting biomass in eucalyptus plantations in north-ern Spain (R2 = 0.75; 4 pulses m−2). LiDAR studies carried outin small areas have shown that the height percentiles are usu-ally closely correlated with biomass (González-Ferreiro et al.2012); in the present study, the elevation percentiles (treeheight) were found to be the most important LiDAR variablesto include in multiple linear regression models for estimatingbiomass (FWand DW) at trial level. The results of a study byZhao et al. (2009) showed that the models can accuratelypredict forest biomass and that the predictive performancewas consistent across a range of scales, with R2 rangingfrom 0.80 to 0.95 across all fitted models. However, in astudy conducted by Van Aardt et al. (2006) in a coniferousplantation, the R2 values for volume (0.66) and abovegroundbiomass (0.59) were low, which was attributed to variability inthe volume. Condés et al. (2013) noted that better fits can beachieved with multiple linear regression models by inclusionof a larger number of variables. This was also observed in thepresent study, and the R2 value increased as the number ofindependent variables that make up the model increased,e.g., in the multiple linear regression, the best result was ob-tained for D20 and the Bjor clone, with six variables includedin the model (R2 = 0.98).

Different studies in conifer plantations worldwide also in-dicate that height can be accurately modeled from medium-and low-density laser data. In the abovementioned study onP. radiata, González-Ferreiro et al. (2012) indicated differ-ences in the model fit for height for different data point den-sities (0.5 and 4 pulses m−2), although the difference in thegoodness of fit was only 8%. Thus, although better fits areobtained by using higher point densities, the extra costs in-volved may not be justified by the final result.

In addition to the importance of reducing the cost of theinventory, the LiDAR information obtained is very useful forforest management purposes. Several studies indicate that theuse of LiDAR data generates more accurate inventories thattraditional inventory methods based on field measurements(Maltamo et al. 2004; Næsset 2002). The combined use ofLiDAR technology and advanced statistical techniques hasled to a number of different studies exploring their potentialfor producing accurate results in forest biomass research (Choet al. 2012; Gleason and Im 2012; Lefsky et al. 2001, 2005;Weishampel et al. 1996; Wulder 1998). In this respect, selec-tion of a suitable statistical approach is essential considering

Annals of Forest Science (2017) 74: 69 Page 13 of 16 69

that the models will be used for predictive purposes. Besidesbeing easy to use and interpret, the non-parametric method(CART) includes classification rules and is thus a very usefultool for LiDAR-based inventories.

As mentioned earlier, the methods used in this study enablemore accurate inventories to be carried out and at a fraction ofthe cost than possible by traditional methods (assuming thetimely availability of suitable data) (Means and Acker 2000).The forest variables estimated from the LiDAR data in thisstudy are of great interest to the timber industry and representinformation that is expensive to collect in the field. The resultsindicate the good relationship between LiDAR data and thevarious forest variables considered. This is of great value inforest management, providing a tool to determine the best timefor harvesting the energy crop, as well as for monitoring andmanaging plots (Lim et al. 2003).

The resultsobtained in this study, in termsof themodels fitted,indicate that forest energy crops can be accuratelymodeled fromlow-intensity laser data and that the models are similar to thosereported in the international scientific literature.

5 Conclusions

The study findings show that low density LiDAR data (0.5pulses m−2) can be used to construct models to estimate themain variables (structural attributes and biomass yield) of in-terest in the management of short-rotation (willow) energycrops. According to the moderate to highly accurate estimatesobtained, the models developed on the basis of LiDAR datacan be used to produce good predictions and estimates for anSRC crop, and would serve as a management tool for improv-ing and optimizing follow-up decisions related to a commer-cial crop. In view of the results obtained, acquisition or pur-chase of low-density LiDAR information to facilitate themon-itoring of the energy plantations can be considered from acommercial point of view. The parametric and non-parametric statistical methods tested in the present study(MLR and CART) provided complementary and robust infor-mation for predicting stand variables in an SRC energy cropfrom low-density LiDAR data (0.5 pulses m−2) and thereforecan be considered suitable for developing models for accurateestimation of forest variables. The predictive power of bothmethods was generally high (particularly when limited datawere available for fitting), althoughMLR models are easier tointerpret and apply.

Acknowledgements This work was funded by the Hunosa Group coalmining company. The authors acknowledge the helpful co-operation ofstaff from Hunosa Group in this study.

Funding information The research was supported by the HunosaChair at the University of Oviedo.

References

Bell JF (1999) Tree-based methods. The use of classification trees topredict species distributions. In: Fielding A (ed) Machine learningmethods for ecological applications. Kluwer, Dordrecht

Belsley DA, Kuh E, Welsch RE (2005) Regression diagnostics: identify-ing influential data and sources of collinearity. Wiley, New York

Breiman L, Friedman JH, Olshen RA, Stone CJ (1984) Classification andregression trees. Wadsworth and Brooks/Cole, Monterey

Broeckx LS, VerlindenMS, Ceulemans R (2012) Establishment and two-year growth of a bio-energy plantation with fast-growing Populustrees in Flanders (Belgium): effects of genotype and former land use.Biomass Bioenergy 42:151–163. https://doi.org/10.1016/j.biombioe.2012.03.005

Cho MA, Mathieu R, Asner GP, Naidoo L, van Aardt J, Ramoelo A,Debba P, Wessels K, Main R, Smit IPJ, Erasmus B (2012)Mapping tree species composition in south African savannas usingan integrated airborne spectral and LiDAR system. Remote SensEnviron 125:214–226. https://doi.org/10.1016/j.rse.2012.07.010

Clark ML, Clark DB, Roberts DA (2004) Small-footprint LIDAR esti-mation of sub-canopy elevation and tree height in a tropical rainforest landscape. Remote Sens Environ 91:68–89. https://doi.org/10.1016/j.rse.2004.02.008

Condés S, Fernández-Landa A, Rodríguez F (2013) Influencia delinventario de campo en el error de muestreo obtenido en uninventario con tecnología Lidar. 6° Congreso Forestal Español(Sociedad Española de Ciencias Forestales). 6CFE01–432

Coops NC, Hilker T, Wulder MA, St-Onge B, Newnham G, Siggins A,Trofymow JT (2007) Estimating canopy structure of Douglas-firforest stands from discrete-return LiDAR. Trees 21:295. https://doi.org/10.1007/s00468-006-0119-6

Dalponte M, Bruzzone L, Gianelle DA (2011) System for the estimationof single-tree stem diameter and volume using multireturn LIDARdata. IEEE Trans Geosci Remote Sens 49:2479–2490. https://doi.org/10.1109/TGRS.2011.2107744

Deconinck E, Hancock T, Coomans D, Massart DJ, Vander Heyden Y(2005) Classification of drugs in absorption classes using the clas-sification and regression trees (CART) methodology. J PharmBiomed Anal 39:91–103. https://doi.org/10.1016/j.jpba.2005.03.008

Edenhofer O, Pichs-Madruga R, Sokona Y, Seyboth K, Arvizu D,Bruckner T, Christensen J, Devernay J-M, Faaij A, Fischedick M,Goldstein G, Hansen J, Huckerby A, Jäger-Waldau S, Kadner D,Kammen V, Krey A, Kumar A, Lewis O, Lucon P, Matschoss L,Maurice C,Mitchell W,Moomaw J,Moreira A, Nadai LJ, Nilsson J,Nyboer A, Rahman J, Sathaye J, Sawin R, Schaeffer T, Schei S,Schlömer R, Sims A, Verbruggen C, Von Stechow K, Urama R,Wiser F, Yamba T, Zwickel (2011) Summary for policy makers.In: Edenhofer O, Pichs-Madruga R, Sokona Y, Seyboth K,Matschoss P, Kadner S, Zwickel T, Eickemeier P, Hansen G,Schlöme S, Stechow CV (eds) IPCC special report on renewableenergy sources and climate change mitigation. CambridgeUniversity Press, Cambridge

EEA (2011) Biogeographical regions. European Environment Agency,Copenhagen [online] URL: http://www.eea.europa.eu/dataand-maps/data/biogeographical-regions-europe-1

Eid T, Gobakken T, Næsset E (2004) Comparing stand inventories forlarge areas based on photo-interpretation and laser scanning bymeans of cost-plus-loss analyses. Scand J For Res 19:512–523

Estornell J, Ruiz LA, Velázquez-Martí B, Fernández-Sarría A (2011)Estimation of shrub biomass by airborne LiDAR data in small foreststands. For Ecol Manag 262:1697–1703. https://doi.org/10.1016/j.foreco.2011.07.026

Estornell J, Velázquez-Martí B, López-Cortés I, Salazar D, Fernández-Sarría A (2014) Estimation of wood volume and height of olive tree

69 Page 14 of 16 Annals of Forest Science (2017) 74: 69

plantations using airborne discrete-return LiDAR data. GISciRemote Sens 51:17–29. https://doi.org/10.1080/15481603.2014.883209

Falkowski MJ, Smith AMS, Hudak AT, Gessler PE, Vierling LA,Crookston NL (2006) Automated estimation of individual conifertree height and crown diameter via two-dimensional spatial waveletanalysis of LiDAR data. Can J Remote Sens 32:153–161. https://doi.org/10.5589/m06-005

Forest Research (2003) Mensurational variables protocol. In: YieldModels for Energy Coppice of Poplar and Willow. 1–14

García M, Riaño D, Chuvieco E, Danson FM (2010) Estimating biomasscarbon stocks for a Mediterranean forest in central Spain usingLiDAR height and intensity data. Remote Sens Environ 114:816–830. https://doi.org/10.1016/j.rse.2009.11.021

García-Gutierrez J, González-Ferreiro E, Riquelme-Santos JC, MirandaD, Dieguez-Aranda U, Navarro-Cerrillo RM (2014) Evolutionaryfeature selection to estimate forest stand variables using LiDAR.Int J Appl Earth Obs Geoinf 26:119–131. https://doi.org/10.1016/j.jag.2013.06.005

Gleason CJ, Im J (2012) Forest biomass estimation from airborne LiDARdata using machine learning approaches. Remote Sens Environ 125:80–91. https://doi.org/10.1016/j.rse.2012.07.006

Gonçalves-Seco L, González-Ferreiro E, Diéguez-Aranda U, Fraga-Bugallo B, Crecente R, Miranda D (2011) Assessing attributes ofhigh densityEucalyptus globulus stands using airborne laser scannerdata. Int J Remote Sens 32:9821–9841. https://doi.org/10.1080/01431161.2011.593583

González-Ferreiro E, Diéguez-Aranda U, Miranda D (2012) Estimationof stand variables in Pinus radiata D. Don plantations using differ-ent LiDAR pulse densities. Forestry 85:281–292. https://doi.org/10.1093/forestry/cps002

González-Ferreiro E, Miranda D, Barreiro-Fernández L, Buján S, García-Gutiérrez J, Dieguez-Aranda U (2013) Modelling stand biomassfractions in Galician Eucalyptus globulus plantations by use of dif-ferent LiDAR pulse densities. For Syst 22:510–525. https://doi.org/10.5424/fs/2013223-03878

Gordon L (2013) Using classification and regression trees (CART) inSAS for applications in public health. SAS Glob Forum 1–8

Graham B (2008) Using low density, small footprint LIDAR in forestinventory applications in the Southeastern U.S

Hall SA, Burke IC, Box DO, Kaufmann MR, Stoker JM (2005)Estimating stand structure using discrete-return Lidar: an examplefrom low density, fire prone ponderosa pine forests. For Ecol Manag208:189–209. https://doi.org/10.1016/j.foreco.2004.12.001

Hawbaker TJ, Gobakken T, Lesak A, Trømborg E, Contrucci K, RadeloffV (2010) Light detection and ranging-based measures of mixedhardwood forest structure. For Sci 56:313–326

Hayashi R, Weiskittel A, Sader S (2014) Assessing the feasibility of low-density LiDAR for stand inventory attribute predictions in complexand managed forests of northern Maine, USA. Forests 5:363–383.https://doi.org/10.3390/f5020363

Hirsch RP (1991) Validation samples. Biometrics 47:1193–1194IBM Corp. Released (2015) IBM SPSS Statistics for Windows, Version

23.0. IBM Corp, ArmonkIDAE (2007) Biomasa: cultivos energéticos. Instituto para la

Diversificación y el Ahorro de la Energía (IDAE), 49 pp, MadridJama A, Nowak W (2012) Willow (Salix viminalis L.) in purifying sew-

age sludge treated soils. Polish J Agron 9:3–6Keoleian GA, Volk TA (2005) Renewable energy from willow biomass

crops: life cycle energy, environmental and economic performance.CRC Crit Rev Plant Sci 24:385–406. https://doi.org/10.1080/07352680500316334

Kozak A, Kozak R (2003) Does cross validation provide additional in-formation in the evaluation of regression models? Can J For Res 33:976–987. https://doi.org/10.1139/X03-022

Kraus K, Mikhail EM (1972) Linear least squares interpolation.Photogramm Eng 38:1016–1029

Kraus K, Pfeifer N (1998) Determination of terrain models in woodedareas with airborne laser scanner data. ISPRS J PhotogrammRemote Sens 53:193–203. https://doi.org/10.1016/S0924-2716(98)00009-4

Kuzovkina YA, Quigley MF (2005) Willows beyond wetlands: uses ofSalix L. species for environmental projects. Water Air Soil Pollut162:183–204. https://doi.org/10.1007/s11270-005-6272-5

Lefsky MA, Cohen W, Spies T (2001) An evaluation of alternate remotesensing products for forest inventory, monitoring, and mapping ofDouglas-fir forests in western Oregon. Can J For Res 31:78–87.https://doi.org/10.1139/cjfr-31-1-78

Lefsky MA, Cohen WB, Parker GG, Harding DJ (2002) LiDAR remotesensing for ecosystem studies: LiDAR, an emerging remote sensingtechnology that directly measures the three-dimensional distributionof plant canopies, can accurately estimate vegetation structural attri-butes and should be of particular interest to forest, landscape, andglobal ecologists. Bioscience 52:19–30. https://doi.org/10.1641/0006-3568

Lefsky MA, Hudak AT, Cohen WB, Acker SA (2005) Geographic vari-ability in LiDAR predictions of forest stand structure in the PacificNorthwest. Remote Sens Environ 95(4):532–548. https://doi.org/10.1016/j.rse.2004.12.022.

Lever J, Krzywinski M, Altman N (2016) Points of significance: modelselection and overfitting. Nat Methods 13:703–704. https://doi.org/10.1038/nmeth.3968

Li Y, Andersen H-E, McGaughey R (2008) A comparison of statisticalmethods for estimating forest biomass from light detection and rang-ing data. J Appl For 23:223–231

Lim K, Treitz P, Wulder M (2003) LiDAR remote sensing of foreststructure. Prog Phys Geogr 1:1–38. https://doi.org/10.1191/0309133303pp360ra

Magdaleno-Mas F, Martínez-Romero R (2006) Aplicaciones de lateledetección láser (LiDAR) en la caracterización y gestión delmedio fluvial. Ingeniería Civil 142:1–15

Magnussen S, Boudewyn P (1998) Derivations of stand heights fromairborne laser scanner data with canopy-based quantile estimators.Can J For Res 28:1016–1031

Maltamo M, Mustonen K, Hyyppä J, Pitkänen J, Yu X (2004) The accu-racy of estimating individual tree variables with airborne laser scan-ning in a boreal nature reserve. Can J For Res 34:1791–1801. https://doi.org/10.1139/X04-055

Mandeville PB (2008) Tema 18: ¿Por qué se deben centrar las covariablesen regresión lineal? Ciencia Uanl 3:300–305

Maune D (2007) Digital Elevation Model Technologies andApplications: The DEM User’s Manual, 2nd edn. AmericanSociety for Photogrammetry and Remote Sensing, Bethesda

McGaughey R (2010) Fusion/LDV: software for LiDAR data analysisand visualization. USDA Forest Service. Pacific NorthwestResearch Station. FUSION Manual. FUSION Version 2.80

Means J, Acker S (2000) Predicting forest stand characteristics with air-borne scanning LiDAR. Photogramm Eng Remote Sens 66:1367–1371. https://doi.org/10.1016/S0034-4257(01)00290-5

Means JE, Acker SA, Harding DJ, Blair JB, Lefsky MA, Cohen WB,Harmon ME, McKee WA (1999) Use of large-footprint scanningairborne LiDAR to estimate forest stand characteristics in the west-ern cascades of Oregon. Remote Sens Environ 67:298–308. https://doi.org/10.1016/S0034-4257(98)00091-1

Montealegre Gracia AL, Lamelas Gracia MT, De la Riva Fernández J,García Martín A, Escribano Bernal F (2015) Cartografía de labiomasa aérea total en masas de Pinus halepensisMill. en el entornode Zaragoza mediante datos LiDAR-PNOAy trabajo de campo. In:Análisis espacial y representación geográfica: innovación yaplicación. pp 769–776

Annals of Forest Science (2017) 74: 69 Page 15 of 16 69

Myers RH (1990) Classical and modern regression with applications, 2ndedn. Duxbury Press, Belmont

Næsset E (1997) Determination of mean tree height of forest stands usingairborne laser scanner data. ISPRS J Photogramm Remote Sens 52:49–56

Næsset E (2002) Predicting forest stand characteristics with airbornescanning laser using a practical two-stage procedure and field data.Remote Sens Environ 80:88–99. https://doi.org/10.1016/S0034-4257(01)00290-5

Næsset E (2011) Estimating above-ground biomass in young forests withairborne laser scanning. Remote Sens 32:473–501. https://doi.org/10.1080/01431160903474970

Næsset E, Gobakken T (2008) Estimation of above and below groundbiomass across regions of the boreal forest zone using airborne laser.Remote Sens Environ 112:3079–3090. https://doi.org/10.1016/j.rse.2008.03.004

Næsset E, Gobakken T, Holmgren J, Hyyppä H, Hyyppä J, Maltamo M,Nilsson M, Olsson H, Persson A, Söderman U (2004) Laser scan-ning of forest resources: the Nordic experience. Scand J For Res 19:482–499. https://doi.org/10.1080/02827580410019553

Paredes-Sánchez JP, García-Elcoro VE, Rosillo-Calle F, Xiberta-Bernat J(2016) Assessment of forest bioenergy potential in a coal-producingarea in Asturias (Spain) and recommendations for setting up aBiomass Logistic Centre (BLC). Appl Energy 171:133–141.https://doi.org/10.1016/j.apenergy.2016.03.009

Rosso L, Facciotto G, Bergante S, Vietto L, Nervo G (2013) Selectionand testing of Populus alba and Salix spp. as bioenergy feedstock:preliminary results. Appl Energy 102:87–92. https://doi.org/10.1016/j.apenergy.2012.07.042

Sánchez Martínez FJ, Andrés-Moreno C, García-Oliva-González M(2011) Introducción a la tecnología LiDAR. In: Marquínez-GarcíaJ, Yagüe-Córdova J (eds) Guía para la aplicación del sistemanacional de cartografía de zonas inundables. Dirección General delAgua. Ministerio de Medio Ambiente y Medio Rural y Marino.Madrid, pp 17–32

Seidel D, Ammer C (2014) Efficient measurements of basal area in shortrotation forests based on terrestrial laser scanning under special con-sideration of shadowing. IForest 7:227–232. https://doi.org/10.3832/ifor1084-007

Song Y, Lu Y (2015) Decision tree methods : applications for classifica-tion and prediction. Shanghai Arch Psychiatry 27:130–135. 10.11919/j.issn.1002-0829.215044

Styles D, Jones M (2007) Energy crops in Ireland: quantifying the poten-tial life-cycle greenhouse gas reductions of energy-crop electricity.Biomass Bioenergy 31:759–772. https://doi.org/10.1016/j.biombioe.2007.05.003

Suárez-Antuña F (2005) La organización de los espacios mineros de lahulla en Asturias. Scripta Nova: Revista Electrónica de Geografía yCiencias Sociales 9:203

Tahvanainen L, Rytko V (1999) Biomass production of Salix viminalis insouthern Finland and the effect of soil properties and climate condi-tions on its production and survival. Biomass Bioenergy 16:103–117. https://doi.org/10.1016/S0961-9534(98)00074-9

Tanarro I (2010) Spectroscopic Techniques. In: Escribano R, Tanarro I(eds) Spectroscopic of the Atmospheres. Consejo Superior deInvestigaciones Científicas (CSIC), Madrid

Thomas V, Treitz P, McCaughey JH, Morrison I (2006) Mapping stand-level forest biophysical variables for a mixedwood boreal forestusing Lidar: an examination of scanning density. Can J For Res36:34–47. https://doi.org/10.1139/x05-230

Timofeev R (2004) Classification and regression trees (CART) theoryand applications. Doctoral dissertation. Humboldt University, Berlin

Treitz P, Kevin L, Murray W, Doug P, Nesbitt D, Etheridge D, (2010)LiDAR data acquisition and processing protocols for forest resourceinventories in Ontario, Canada. In: Silvilaser 2010: The 10thInternational Conference on LiDAR Applications for AssessingForest Ecosystems, Freiburg, Germany

Van Aardt JA, Wynne RH, Oderwald RG (2006) Forest volume andbiomass estimation using small-footprint LiDAR-distributional pa-rameters on a per-segment basis. For Sci 52:636–649

Venturi P, Gigler JK, HuismanW (1999) Economical and technical com-parison between herbaceous (Miscanthus x giganteus) and woodyenergy crops (Salix viminalis). Renew Energy 16:1023–1026.https://doi.org/10.1016/S0960-1481(98)00363-2

Wehr A, Lohr U (1999) Airborne laser scanning an introduction andoverview. ISPRS J Photogramm Remote Sens 54:68–82

Weih M (2004) Intensive short rotation forestry in boreal climates: pres-ent and future perspectives. Can J For Res 34:1369–1378. https://doi.org/10.1139/x04-090

Weishampel JF, Ranson KJ, Harding DJ (1996) Remote sensing of forestcanopies. Selbyana 17:6–14

Wulder M (1998) Optical remote-sensing technique for the assessment offorest inventory and biophysical parameters. Prog Phys Geogr 22:449–476. https://doi.org/10.1177/030913339802200402

Zhao K, Popescu S, Nelson R (2009) LiDAR remote sensing of forestbiomass: a scale-invariant estimation approach using airborne lasers.Remote Sens Environ 113:182–196. https://doi.org/10.1016/j.rse.2008.09.009

Zimble DA, Evans DL, Carlson GC, Parker RC, Grado SC, Gerard PD(2003) Characterizing vertical forest structure using small-footprintairborne LiDAR. Remote Sens Environ 87:171–182. https://doi.org/10.1016/S0034-4257(03)00139-1

Zurba K, Oertel C, Matschullat J (2013) CO2 emissions from willow andpoplar short rotation forestry (SRF) on a derelict mining soil. 1–4

69 Page 16 of 16 Annals of Forest Science (2017) 74: 69